GraphSAGE, introduced by William Hamilton, Rex Ying, and Jure Leskovec in the paper “Inductive Representation Learning on Large Graphs” (submitted to arXiv on June 7, 2017), addressed a basic limitation of earlier graph embedding methods. Techniques like node2vec and the original spectral graph networks were transductive: they learned a fixed embedding for every node in a single graph and could not produce embeddings for nodes that appeared later. GraphSAGE made the process inductive.
Instead of memorizing one vector per node, GraphSAGE learns a set of aggregator functions that sample a fixed number of neighbors and combine their features. To embed a node, the model gathers features from its sampled neighborhood, aggregates them (by mean, pooling, or an LSTM), and repeats this over a few hops. Because the learned functions depend only on local feature patterns, they can be applied to brand new nodes or even entirely new graphs.
The authors demonstrated the method on citation networks, a large Reddit post graph, and protein-protein interaction data, showing strong node classification performance and generalization to unseen data. The sampling strategy also kept memory and compute manageable on graphs too large to process whole.
This matters in practice because real systems are constantly adding new users, products, and transactions. GraphSAGE made it feasible to embed those new entities on the fly without retraining the whole model, which is why it became a foundation for industrial graph machine learning.